A study of classification techniques

Abstract

The following thesis studies both parametric and non parametric approaches to classification. Among the various methods which exist, three are developed in detail. They are the Bayes rule, classification based on kernel density estimation and Fisher's discriminant function applied to the ranked data. Furthermore, we propose a new classification rule based on ranks. A Monte Carlo simulation study is then performed to test this new method and compare it with the other three classification techniques. The simulations indicate that the new classification rule performs well in many cases and that it is most effective when the number of dimensions are high and few observations are available. In this particular situation, the new classification rule proposed had the lowest probability of misclassification.